Initial commit: Llama 4 Scout BI-V150 deployment bundle

2026-10-23 13:41:41 +08:00
commit 7ce7159d36
6 changed files with 296 additions and 0 deletions
--- a/vllm_llama4_chunked_attn.patch
+++ b/vllm_llama4_chunked_attn.patch
@@ -0,0 +1,16 @@
+--- a/vllm/v1/attention/backends/utils.py
+++ b/vllm/v1/attention/backends/utils.py
+@@ -346,7 +346,13 @@
+     return CommonAttentionMetadata(
+         query_start_loc_cpu=query_start_loc_cpu,
+         query_start_loc=query_start_loc_cpu.to(device=device, non_blocking=True),
+        key_start_loc=torch.from_numpy(
+            np.concatenate(
+                (np.zeros(1, dtype=np.int32), np.cumsum(seqlens_k_local, dtype=np.int32))
+            )
+        ).to(device=device, non_blocking=True),
+         seq_lens=seq_lens_cpu.to(device=device, non_blocking=True),
+        seq_lens_np=seqlens_k_local,
+         num_reqs=len(seq_lens_cpu),
+         num_actual_tokens=common_attn_metadata.num_actual_tokens,
+         max_query_len=seqlens_q_local.max(),